Investigation of Co-training Views and Variations for Semantic Role Labeling

Co-training, as a semi-supervised learning method, has been recently applied to semantic role labeling to reduce the need for costly annotated data using unannotated data. A main concern in co-training is how to split the problem into multiple views to derive learning features, so that they can effectively train each other. We investigate various feature splits based on two SRL views, constituency and dependency, with different variations of the algorithm. Balancing the feature split in terms of the performance of the underlying classifiers showed to be useful. Also, co-training with a common training set performed better than when separate training sets are used for co-trained classifiers.